E167: Nvidia smashes earnings (again), Google's Woke AI disaster, Groq's LPU breakthrough & more

00:00:00.000 | All right, everybody, welcome back to your favorite podcast

00:00:02.760 | of all time, the all in podcast, episode 160 something with me

00:00:07.680 | again, Chamath Palihapitiya. He's the CEO of a company and

00:00:12.160 | invest in startups and his firm is called social capital. We

00:00:16.960 | also have David Freeberg, the sultan of science. He's now a

00:00:20.000 | CEO as well. And we have David Sachs from craft ventures in

00:00:25.640 | some undisclosed hotel room somewhere. How are we doing

00:00:28.520 | boys? Good. Thank you. This is not hard. Your intro be any more

00:00:33.320 | low energy and dragged out. I'm sick. What do you want me to

00:00:37.600 | do? Effort. All right, here we go. Give me one more shot. Watch

00:00:41.200 | this. Watch this watch professional. You want

00:00:43.400 | professionalism? The effort Come on. Here we go. You want

00:00:46.120 | professionalism? I'll show you guys professionalism. Is that

00:00:48.560 | binaka? What was that? That's not good. Oh, it is the secret.

00:00:58.560 | Rain Man, David.

00:00:59.600 | All right, everybody, welcome to the all in podcast episode

00:01:13.880 | 167 168. With me, of course, the rain man himself, David Sachs,

00:01:18.320 | the dictator, Chairman Chabot, Polly hoppity, and our sultan of

00:01:22.120 | science, David Freeberg. How we doing, boys? Great. How are you?

00:01:24.440 | Is it 167 or 168? No, who cares? We at least get you to know the

00:01:30.040 | episode number. Who cares? Unfortunately, or fortunately,

00:01:34.120 | we're going to be doing this thing forever. The audience

00:01:36.240 | demands it. It doesn't matter. This is like a Twilight Zone

00:01:39.200 | episode. We're going to be trapped in these four bubbles

00:01:41.800 | forever. Superman. It's a it is it's this is like the it is the

00:01:46.560 | gift. glass. Was that? Yeah, Neil before Zod. And he spun

00:01:53.680 | through the universe and the plastic thing forever for for

00:01:56.720 | infinity until that until Superman took the nuclear bomb

00:02:00.640 | out of the Eiffel Tower and threw it into space and blew it

00:02:04.840 | up. And for you know, my background today, I think I'm

00:02:07.360 | gonna have to change now that you've referenced this

00:02:09.280 | important scene. That was the best moment of that movie, J.

00:02:11.760 | Cal, where Terrence stamps says Neil to the President and the

00:02:16.000 | President says, Oh, God, yes. And Terrence was Zod.

00:02:20.520 | Neil before Zod.

00:02:25.320 | Superman two is pretty much the best. Yeah, you know, like

00:02:29.600 | Empire Strikes Back like Terminator two, it's always the

00:02:32.440 | second one. That's the best one. All right, everybody, we got a

00:02:34.840 | lot to talk about today. Apologies for my voice a little

00:02:37.240 | bit of a cold Nvidia blew the doors off their earnings for the

00:02:40.120 | third straight quarter. shares were up 15% on Thursday

00:02:44.080 | representing a nearly $250 billion jump in market cap. So

00:02:49.000 | let's just let that sit in for a second. This is the largest

00:02:52.800 | single day gain in market cap. $247 billion added in market

00:02:59.000 | cap previously meta did something similar earlier this

00:03:02.120 | year. Remember everybody was down on that stock because they

00:03:04.400 | were doing all the crazy stuff with reality labs and then they

00:03:07.640 | got focused and laid off 20,000 people. They added $196 billion.

00:03:11.880 | In other words, they added like two and a half Airbnbs to their

00:03:15.840 | valuation. But let's just get to the results. The results are

00:03:18.000 | absolutely stunning. I dare I say unprecedented Q four revenue

00:03:21.600 | 22.1 billion that's up 22% quarter of a quarter of 265%

00:03:27.520 | year over year. The net income was 12.3 billion nine x year over

00:03:32.560 | year, and the gross margin of 76% was up two points quarter of a

00:03:36.680 | quarter 12.7% year over year. But look at this revenue ramp.

00:03:40.840 | This is extraordinary. q1 of 2024 this juggernaut starts and

00:03:47.680 | it does not stop and it doesn't look like it's going to stop

00:03:50.480 | just to run up from 7 billion all the way to 22 billion in

00:03:54.760 | revenue for the quarter. Absolutely extraordinary. And if

00:03:59.440 | you want to know why this is happening, why is Nvidia putting

00:04:02.080 | up these kind of numbers? This chart explains everything. This

00:04:05.280 | is all about data centers. Obviously, if you heard of

00:04:08.760 | Nvidia before the AI boom, it was gaming, professional

00:04:12.440 | visualizations, you know, I think people making movies and

00:04:14.960 | stuff like that autos used Nvidia for self driving, that

00:04:18.560 | kind of stuff. But if you look at this chart, you'll see data

00:04:21.080 | centers, just starting four quarters ago, starts to ramp up

00:04:26.160 | as everybody builds out the infrastructure for new data

00:04:30.760 | centers to deal with generative AI.

00:04:32.760 | So just to add one point here, Jason, so what you can see is

00:04:36.960 | that Nvidia was around for a long time, and it was making

00:04:40.320 | these chips, these GPUs, as opposed to CPUs. And they were

00:04:44.720 | primarily used by games, and by virtual reality software,

00:04:51.040 | because GPUs are better, obviously, at graphical

00:04:54.080 | processing, they use vector math to create these like 3d

00:04:57.160 | worlds. And this vector math that they use to create these

00:05:02.040 | 3d worlds is also the same vector math that AI uses to

00:05:05.920 | reach its outcomes. So with the explosion of LLM, it turns out

00:05:10.760 | that these GPUs are the right chips that you need for these

00:05:13.920 | cloud service providers to build out these big data centers to

00:05:18.040 | serve now all of these new AI applications. So Nvidia was in

00:05:22.520 | the perfect place at the perfect time. And that's why it's just

00:05:26.240 | exploded. And what you're seeing is the build out of this new

00:05:31.560 | cloud service infrastructure for AI.

00:05:33.720 | Yeah. And it also helping the stock is the fact that they

00:05:38.360 | bought back 2.7 billion worth of their shares as part of a $25

00:05:41.280 | billion buyback plan. But this company is firing on all

00:05:44.520 | cylinders revenues, obviously ripping as people put in orders

00:05:47.680 | to replace all of the data centers out there, or at least

00:05:51.520 | augment them with this technology with GPUs, a 100s,

00:05:55.200 | H 100s, etc. The gross margins been expanding, they have huge

00:05:59.360 | profits, and they're still projecting more growth in q1,

00:06:02.960 | around 24 billion, which would be a 3x increase year over year.

00:06:07.040 | And this obviously has made the entire market rip. As Nvidia

00:06:11.000 | goes, so does the market right now. And the s&p 500 and Nasdaq

00:06:15.200 | are at record highs at the time of this taping Chamath your

00:06:19.560 | general thoughts here on something I don't think anybody

00:06:23.160 | saw coming. Except for you and your investment in grok maybe,

00:06:28.080 | and a couple of others. I think what I would tell you is that

00:06:30.800 | the bigger principle, and we've talked about this a lot, Jason,

00:06:35.120 | is that in capitalism, when you over earn for enough of a time,

00:06:41.600 | what happens is competitors decide to try to compete away

00:06:44.480 | your earnings. In the absence of a monopoly, the amount of time

00:06:48.360 | that you have tends to be small, and it shrinks. So in the case

00:06:52.080 | of a monopoly, for example, take Google, you can over earn for

00:06:55.480 | decades. And it takes a very, very long time for somebody to

00:06:59.440 | try to displace you, we're just starting to see the beginnings

00:07:02.520 | of that, with things like perplexity and other services

00:07:06.280 | that are chipping away at the Google monopoly. But at some

00:07:10.320 | point in time, all of these excess profits are competed

00:07:14.240 | away. In the case of Nvidia, what you're now starting to see

00:07:18.440 | is them over earn in a very massive way. So the real

00:07:22.520 | question is who will step up to try to compete away those

00:07:27.200 | profits. The old Bezos quote, right, your margin is my

00:07:31.680 | opportunity. And I think we're starting to see and you've

00:07:34.360 | mentioned grok, who had a super viral moment, I think this week.

00:07:37.800 | But you're starting to see the emergence of a more detailed

00:07:42.000 | understanding of what this market actually means. And as a

00:07:45.320 | result, who will compete away the inference market who will

00:07:49.440 | compete away the training market, and the economics of

00:07:52.240 | that are just becoming known to now more and more people.

00:07:54.720 | Freeburg, your thoughts, we were talking, I think, was last week

00:07:58.400 | or the week before about possibility of Nvidia being a

00:08:01.680 | $10 trillion company, largest company in the world, what are

00:08:04.120 | your thoughts on the spectacular results? And then, to Matt's

00:08:07.680 | point, everybody is watching this going, Hmm, maybe I can get

00:08:10.920 | a slice of that pie. And maybe I can create a more competitive

00:08:15.000 | offering. Obviously, we saw Sam Altman, rumored to be raising

00:08:18.880 | 7 trillion, which feels like a fake number of feels like that's

00:08:21.480 | maybe the market size or something. But your thoughts,

00:08:23.560 | you

00:08:23.720 | don't think anything's changed on the Nvidia front, there's this

00:08:26.160 | accelerated compute build out underway. In data centers,

00:08:30.320 | everyone's building infrastructure, and then

00:08:32.200 | everyone's trying to build applications and tools and

00:08:35.000 | services on top of that infrastructure. The

00:08:37.280 | infrastructure build out is kind of the first phase. The real

00:08:40.200 | question ultimately will be, does the initial cost of the

00:08:43.320 | infrastructure exceed the ultimate value that's going to

00:08:46.640 | be realized on the application layer? In the early days of the

00:08:50.200 | internet, a lot of people were buying Oracle servers, they were

00:08:55.000 | like 3000 bucks a server. And they were running these Oracle

00:08:58.960 | servers out of an internet connected data center. And it,

00:09:02.160 | you know, took a couple of years before folks realized that for

00:09:05.400 | large scale distributed compute applications, you're better off

00:09:10.360 | using cheaper hardware, you know, cheaper server racks,

00:09:13.760 | cheaper hard drives, cheaper buses, and assuming a shorter

00:09:17.800 | lifespan on those servers, and you can cycle them in and out.

00:09:21.240 | And you didn't need the redundancy, you didn't need the

00:09:23.480 | certainty, you didn't need the runtime guarantees. And so you

00:09:27.240 | could use a lower cost, higher failure rate, but much, much net

00:09:32.680 | lower cost kind of approach to building out a data center for

00:09:35.960 | internet serving. And so the Oracle servers didn't really

00:09:39.360 | take the market. And early on, everyone thought that they would

00:09:41.960 | so I think to my point is right now, Nvidia has been at this for

00:09:44.960 | a very long time. And the real question is how much of an

00:09:48.520 | advantage do they have, particularly that there is this

00:09:51.480 | need to use fabs to build replacement technology. So over

00:09:54.960 | time, will there be better solutions that use hardware

00:09:57.200 | that's not as good, but the software figures out and they

00:09:59.360 | build new architecture for running on that hardware in a

00:10:02.000 | way that kind of mimics what we saw in the early days of the

00:10:04.320 | build out of the internet. So TBD, right? The same is true in

00:10:09.360 | switches, right? So in networking, a lot of the high

00:10:12.320 | end high quality networking companies got beaten up when

00:10:16.520 | lower cost solutions came to market later. And so they look

00:10:20.040 | like they were going to be the biggest business ever. I mean,

00:10:21.640 | you can look at Cisco, during the early days of the internet

00:10:24.400 | build out, and everyone thought Cisco was the picks and shovels

00:10:27.880 | of the internet, they were going to make all the all the values

00:10:29.920 | going to go to Cisco. So we're kind of in that same phase right

00:10:32.960 | now with Nvidia, the real question is, is this going to be

00:10:36.040 | a much harder hill to compete on than we've ever seen, given the

00:10:39.880 | development cycle on chips and the requirement to use these

00:10:42.400 | fabs to build chips. It may be a harder hill to kind of get up

00:10:45.640 | sex. So we'll see your thoughts. You think we're getting to the

00:10:48.200 | point where maybe we'll have bought too many of these built

00:10:51.840 | out too much infrastructure, and it will take time for the

00:10:54.160 | application layer as freeberg was alluding to to monetize it?

00:10:58.360 | Well, I think the question everyone's asking right now is,

00:11:00.680 | are these results sustainable? Can Nvidia keep growing at these

00:11:05.400 | astounding rates? You know, will the build out continue and the

00:11:09.200 | comparison everyone's making is to Cisco. And there's this chart

00:11:12.520 | that's been going around, overlaying the Nvidia stock

00:11:16.040 | price on the Cisco stock price. And you can see here, the orange

00:11:19.360 | line is Nvidia and the blue line is Cisco. And it's almost like a

00:11:24.760 | perfect match. Now, what happened is that at a similar

00:11:28.600 | point, in the original build out of the internet of the dot com

00:11:32.800 | era, you had the market crash at the end of March of 2000. And

00:11:39.080 | Cisco never really recovered from that peak valuation. But I

00:11:43.080 | think there's a lot of reasons to believe Nvidia is different.

00:11:45.600 | One is that if you look at Nvidia is multiples are nowhere

00:11:49.080 | near where Cisco's were back then. So the market in 1999 and

00:11:53.160 | early 2000 was way more bubbly than it is now. So Nvidia is

00:11:57.880 | valuation, it's much more grounded in real revenue, real

00:12:01.680 | margins, real profit. Second, you have the issue of

00:12:06.600 | competitive moat. Cisco was selling servers and networking

00:12:11.280 | equipment. Fundamentally, that equipment was much easier to

00:12:15.280 | copy and commoditize than GPUs. These GPU chips are really

00:12:20.800 | complicated. I think Jensen made the point that their hopper 100

00:12:25.800 | product, he said, you know, don't even think of it just like

00:12:29.600 | a chip, there's actually 35,000 components in this product, and

00:12:33.280 | it weighs 70 pounds. This is more like mainframe computer or

00:12:37.160 | something that's dedicated to process somewhere between a rack

00:12:39.960 | server and the entire rack. Yeah, it's giant, and it's heavy,

00:12:44.440 | and it's complex. It does say something here, Chamath, I think

00:12:47.960 | about how well positioned big tech is in terms of seeing an

00:12:55.240 | opportunity, and quickly mobilizing to capture that

00:12:59.440 | opportunity. These servers are being bought by, you know,

00:13:04.120 | people like Amazon, I'm sure Apple, obviously, Facebook meta.

00:13:07.920 | I don't know if Google is buying them as well, I would assume so

00:13:11.000 | Tesla. So everybody's buying these things, and they had tons

00:13:14.960 | of cash sitting around. It is pretty amazing how nimble the

00:13:17.840 | industry is. And this opportunity feels like everybody

00:13:21.520 | is looking at it like mobile and cloud, I have to get mobilized

00:13:24.520 | quickly to not get disrupted.

00:13:26.840 | You're bringing up an excellent point. And I would like to tie

00:13:30.240 | it together with Friedberg's point. So at some point, all of

00:13:34.360 | this spend has to make money, right? Otherwise, you're going

00:13:38.120 | to look really foolish for having spent 20 and 30 and $40

00:13:40.920 | billion. So Nick, if you just go back to the to the revenue slide

00:13:45.080 | of Nvidia, I can try to give you a framing of this at least the

00:13:48.840 | way that I think about it. So if you look at this, like what

00:13:51.320 | you're talking about is look, who is going to spend $22.1

00:13:55.480 | billion? Well, you said it, Jason, it's all a big tech. Why?

00:13:58.960 | Because they have that money on the balance sheet sitting idle.

00:14:01.720 | But when you spend $22 billion, their investors are going to

00:14:07.080 | demand a rate of return on that. And so if you think about what a

00:14:10.160 | reasonable rate of return is, call it 30 40 50%. And then you

00:14:14.000 | factor in and that's profit. And then you factor in all of the

00:14:17.800 | other things that need to support that, that $22 billion

00:14:21.720 | of spend needs to generate probably $45 billion of revenue.

00:14:26.080 | And so Jason, the question to your point, and to Friedberg's

00:14:30.000 | point, the $64,000 question is, who in this last quarter is

00:14:34.840 | going to make 45 billion on that 22 billion of spend. And again,

00:14:39.480 | what I would tell you to be really honest about this is that

00:14:42.080 | what you're seeing is more about big companies, muscling people

00:14:47.400 | around with their balance sheet, and being able to go to Nvidia

00:14:51.320 | and say, I will give you committed pre purchases over the

00:14:55.120 | next three or four quarters. And less about here is a product

00:15:00.000 | that I'm shipping that actually makes money, which I need

00:15:03.400 | enormous more compute resources for. It's not the latter. Most

00:15:09.240 | of the apps, the overwhelming majority of the apps that we're

00:15:12.480 | seeing in AI today, are toy apps that are run as proofs of

00:15:18.040 | concept, and demos, and run in a sandbox. It is not production

00:15:23.680 | code. This is not, we've rebuilt the entire autopilot system for

00:15:30.400 | the Boeing. And it's now run with agents, and bots and all of

00:15:36.160 | this training. That's not what's happening. So it is a really

00:15:40.160 | important question. Today, the demand is clear. It's the big

00:15:43.200 | guys with huge gobs of money. And by the way, Nvidia is super

00:15:47.160 | smart to take it because they can now forecast demand for the

00:15:50.280 | next two or three quarters. I think we still need to see the

00:15:53.800 | next big thing. And if you look in the past, what the past has

00:15:57.120 | showed you, it's the big guys don't really invent the new

00:15:59.400 | things that make a ton of money. It's the new guys, who because

00:16:02.520 | they don't have a lot of money, and they have to be a little bit

00:16:05.560 | more industrious, come up with something really authentic and

00:16:09.040 | new. Yeah, constraint makes for great art. Yeah, we haven't

00:16:11.760 | seen that yet. So I think the revenue scale will continue for

00:16:15.080 | like the next two or three years, probably for Nvidia. But

00:16:19.720 | the real question is, what is the terminal value? And it's the

00:16:22.800 | same thing that SAC showed in that Cisco slide, people

00:16:26.800 | ultimately realized that the value was going to go to other

00:16:32.120 | parts of the stack, the application layer. And as more

00:16:37.040 | and more money was accrued at the application layer of the

00:16:39.360 | internet, less and less revenue multiple and credit was given to

00:16:43.280 | Cisco. And that's nothing against Cisco, because their

00:16:45.520 | revenue continue to compound. Right. And they did an

00:16:48.760 | incredible job, but the valuation got so freeberg. If

00:16:52.400 | we're looking at this chart, the winner of Netflix, the winner of

00:16:56.280 | the Cisco chart might in fact be somebody like Netflix, they

00:16:58.680 | actually got, you know, hundreds of millions of consumers to give

00:17:01.920 | them on Facebook. And then you have Google and Facebook as

00:17:04.440 | well, generating all that traffic. And then YouTube, of

00:17:07.120 | course, who do you see the winner here as in terms of the

00:17:10.840 | application layer? Who are the billion customers here who are

00:17:14.400 | going to spend 20 bucks a month, five bucks a month, whatever it

00:17:16.920 | is, secure?

00:17:17.800 | Well, I mean, let me just start with this important point. If

00:17:20.280 | you look at where that revenue is coming from, to chamat's

00:17:23.360 | point, it's coming from big cloud service providers. So

00:17:28.680 | Google, and others are building out clouds, that other

00:17:34.280 | application developers can build their AI tools and applications

00:17:37.760 | on top of. So a lot of the build out is in these cloud data

00:17:40.960 | centers that are owned and operated by these big tech

00:17:45.600 | companies. The 18 billion of data center revenue that Nvidia

00:17:49.640 | realized is revenue to them. But it's not an operating expense to

00:17:54.400 | the companies that are building out. So this is an important

00:17:57.160 | point on why this is happening at such an accelerated pace.

00:18:00.120 | When a big company buys these chips from Nvidia, they don't

00:18:04.360 | have to from an accounting basis market as an expense in their

00:18:07.480 | income statement, it actually gets booked as a capital

00:18:10.320 | expenditure. In the cash flow statement, it gets put on the

00:18:14.040 | balance sheet, and they depreciate it over time. And so

00:18:17.440 | they can spend $20 billion of cash because Google and others

00:18:20.320 | have 100 billion of cash sitting on the balance sheet. And

00:18:23.320 | they've been struggling to find ways to grow their business

00:18:26.000 | through acquisitions. One of the reasons is they there aren't

00:18:29.840 | enough companies out there that they can buy it a good multiple

00:18:32.280 | that can give them a good increase in profit. The other

00:18:35.120 | one is that antitrust authorities are blocking all of

00:18:37.120 | their acquisitions. And so what do you do with all that cash?

00:18:39.840 | Well, you can build out the next gen of cloud infrastructure, and

00:18:43.680 | you don't have to take the hit on your P&L by doing it. So it

00:18:46.880 | ends up in the balance sheet, and then you depreciate it over

00:18:49.480 | typically four to seven years. So that money gets paid out on

00:18:53.320 | the on the income statement at these big companies over a seven

00:18:56.320 | year period. So there's a really great accounting and M&A

00:19:01.240 | environment driver here that's causing the big cloud data

00:19:05.040 | center providers to step in and say, this is a great time for us

00:19:07.920 | to build out the next generation of infrastructure that could

00:19:11.400 | generate profits for us in the future, because we've got all

00:19:13.720 | this cash sitting around, we don't have to take a P&L hit, we

00:19:16.200 | don't have to acquire a cash burning business. And, you

00:19:20.000 | know, frankly, we're not going to be able to grow through M&A

00:19:21.640 | because of antitrust right now anyway. So there's a lot of

00:19:24.240 | other motivating factors that are causing this near term

00:19:26.480 | acceleration, as they're trying to find ways to grow. Yeah. And

00:19:30.040 | all I know that was an accounting point, but I think

00:19:31.960 | it's a really important one. If you if 100 billion gets

00:19:34.960 | spent this year, you're divided by 425 billion in revenue would

00:19:38.040 | have to come from that or something in that range. Yeah.

00:19:40.120 | And so sacks any guesses,

00:19:42.120 | Jeff, to just keep in mind, I think Fribourg, what you said is

00:19:44.520 | very true for GCP spend, but not necessarily for Google spend.

00:19:49.000 | It's true for AWS spend, but not necessarily for Amazon spend.

00:19:53.040 | And it's true for Azure spend, not true for Microsoft spend.

00:19:56.680 | And it's largely not true for Tesla and Facebook because they

00:19:59.600 | don't have clouds. So I think the question to your point, that

00:20:03.600 | have been, for obvious reasons, Nvidia doesn't disclose it is

00:20:07.120 | what is the percentage of that 21 billion that just went to

00:20:10.600 | those cloud providers that they'll then capitalized to to

00:20:14.600 | everybody else versus what was just absorbed because at

00:20:17.400 | Facebook, Mark had that video about how many h 100. That's all

00:20:20.120 | for him.

00:20:20.600 | Right, but it is still it is still capitalized, is my point.

00:20:24.600 | So they don't have to book that as an expense. It sits on the

00:20:28.080 | balance sheet. Yeah, sure. And they earn it down over time.

00:20:30.760 | You're helping to explain why these big cloud service

00:20:32.760 | providers are spending so much on the

00:20:34.560 | because they're very profitable, and there's nowhere else to put

00:20:37.360 | the money.

00:20:37.760 | Right? Well, so that would seem to indicate that this is more in

00:20:41.960 | the category of one time build out than sustainable ongoing

00:20:45.800 | revenue. I think the big question is the one that mouth

00:20:49.200 | asked, which is, what's the terminal value of Nvidia? I

00:20:52.360 | think, like a simple framework for thinking about that is what

00:20:56.040 | is the total addressable market or tam related to GPUs? And then

00:21:00.280 | what is their market share going to be? Right now, their market

00:21:04.000 | share is something like 91%. That's clearly going to come

00:21:06.640 | down, but their moat appears to be substantial. The Wall Street

00:21:10.600 | analysts I've been listening to think that in five years,

00:21:13.640 | they're still going to have 60 something percent market share.

00:21:16.400 | So they're going to have a substantial percentage of this

00:21:19.320 | market or this Tam, then the question is, I think, with

00:21:22.520 | respect to Tam is, what is one time build out versus steady

00:21:26.680 | state? Now, I think that clearly, there's a lot of build

00:21:32.320 | out happening now, that's almost like a backfill of capacity that

00:21:35.320 | people are realizing they need. But even the numbers you're

00:21:38.360 | seeing this quarter, kind of understated, because, first of

00:21:41.960 | all, Nvidia was supply constrained, they cannot produce

00:21:45.000 | enough chips to satisfy all the demand, their revenue would

00:21:48.360 | would have been even higher, if they had more capacity. Second,

00:21:54.120 | you just look at their forecast. So the fiscal year that just

00:21:57.400 | ended, they did around $60 billion of revenue, they're

00:22:00.440 | forecasting $110 billion for the fiscal year that just started.

00:22:03.920 | So they're already projecting to almost double based on the

00:22:07.680 | demand that they clearly have visibility into already. So it's

00:22:11.440 | very hard to know exactly what the terminal or steady state

00:22:15.000 | value of this market is going to be. Even once the cloud service

00:22:19.920 | providers do this big build out, presumably, there's always going

00:22:22.760 | to be a need to stay up to date with the latest chips, right?

00:22:26.720 | Here's a framework for you, Sax, tell me if this makes sense.

00:22:29.600 | Intel was the basically the mother of all of modern compute

00:22:35.360 | up until today, right? I think the CPU was the the most

00:22:40.120 | fundamental workhorse that enabled local PCs, it enabled

00:22:44.240 | networking, it enabled the internet. And so when you look

00:22:49.320 | at the market cap of it, as an example, that's about 180 odd

00:22:53.720 | billion dollars today. The economy that it created, that it

00:22:59.800 | supports is probably measured, call it in a trillion or $2

00:23:03.360 | trillion, maybe 5 trillion, let's just be really generous,

00:23:06.360 | right. And so you can see that there's this ratio of the

00:23:10.320 | enabler of an economy, and the size of the economy. And those

00:23:15.480 | things tend to be relatively fixed, and they recur

00:23:18.240 | repeatedly over and over and over. If you look at Microsoft,

00:23:21.200 | its market cap relative to the economy that it enables. So the

00:23:24.840 | question for Nvidia, in my mind would be not that it is it not

00:23:28.800 | going to go up in the next 18 to 24 months probably is, for

00:23:32.440 | exactly the reason you said it is super set up to have a very

00:23:35.600 | good meat and beet guidance for the street, which they'll eat

00:23:38.840 | up and all of the algorithms that trade the press releases

00:23:42.160 | will drive the price higher and all of this stuff will just

00:23:44.760 | create a trend upward. I think the bigger question is, if it's

00:23:49.640 | a four or $5 trillion market cap in the next two or three years,

00:23:54.720 | will it support $100 trillion economy? Because that's what you

00:24:01.480 | would need to believe for those ratios to hold. Otherwise,

00:24:04.000 | everything is just broken on the internet.

00:24:05.840 | Yeah, I mean, so the history of the internet is that if you

00:24:09.400 | build it, they will come, meaning that if you make the

00:24:12.200 | investment in the capital assets necessary to power the next

00:24:16.800 | generation of applications, those applications have always

00:24:19.400 | eventually gotten written, even though it was hard to predict

00:24:23.120 | them at the time. So in the late 90s, when we had the whole dot

00:24:26.200 | com bubble, and then bust, you had this tremendous build out

00:24:29.200 | not just of kind of servers and all the networking equipment,

00:24:32.160 | but there was a huge fiber build out. Yep, by all the telecom

00:24:35.120 | companies, and the telecom companies had a Cisco like, you

00:24:38.960 | know, peak, it was worse.

00:24:40.480 | You know, all common them, or they went bankrupt a lot of

00:24:42.680 | them. Yeah, well, the problem there was that a lot of the

00:24:45.520 | buildout happened with debt. And so when you had the dot com

00:24:48.160 | crash, and all the valuations came down to earth, that's why a

00:24:51.880 | lot of them went under. Yeah, Cisco wasn't in that position.

00:24:54.920 | But anyway, my point is, in the early 2000s, when the dot com

00:24:58.600 | crash happened, everyone thought that these telecom companies had

00:25:01.480 | over invested in fiber. As it turns out, all that fiber

00:25:04.800 | eventually got used. The internet went from, you know,

00:25:09.040 | dial up to broadband, we started doing seeing streaming, social

00:25:12.720 | networking, all these applications started eating up

00:25:15.160 | that bandwidth. So I think that the history of these things is

00:25:19.920 | that the applications eventually get written, they get developed

00:25:24.520 | if you build the infrastructure to power them. And I think with

00:25:27.000 | AI, the thing that's exciting to me as someone who's really more

00:25:30.640 | of an application investor, is that we're just at the

00:25:33.200 | beginning, I think of a huge wave of a lot of new creativity

00:25:39.240 | and applications that's going to be written. And it's not just

00:25:41.760 | B2C, it's going to be B2B as well. You guys haven't really

00:25:44.520 | mentioned that it's not just consumers and consumer

00:25:47.280 | applications are going to use these cloud data centers that

00:25:51.720 | are buying up all these GPUs is it's going to be enterprises

00:25:54.240 | too. I mean, these enterprises are using Azure, they're using

00:25:57.600 | Google Cloud, and so forth. So there's a lot I think, that's

00:26:02.600 | still to come. I mean, we're just at the beginning of a wave

00:26:05.320 | that's probably going to last at least a decade.

00:26:07.520 | Yeah. And to your point, one of the reasons YouTube, Google

00:26:12.040 | Photos, iPhoto, a lot of these things happened was because the

00:26:17.200 | infrastructure build out was so great during the dotcom boom,

00:26:20.160 | that the prices for storage, the prices for bandwidth sacks,

00:26:23.360 | plummeted. And then people like Chad Hurley, looked at him like,

00:26:27.480 | you know what, instead of charging people to put a video

00:26:30.160 | on the internet, and then charging them for the bandwidth

00:26:32.280 | they used, we'll just let them upload this stuff to YouTube,

00:26:35.080 | and we'll figure it out later. Same thing with Netflix.

00:26:37.680 | I mean, look, when we were developing PayPal in the late

00:26:40.640 | 90s, really around 1999, you could barely upload a photo to

00:26:46.320 | the internet. I mean, so like the idea of having an account

00:26:49.200 | with a profile photo on it was sort of like, why would you do

00:26:51.560 | that? It's just prohibitively slow, everyone's going to drop

00:26:53.960 | off. Yeah, by 2003, it was fast enough that you could do that.

00:26:59.040 | And that's why social networking happened. I mean, literally,

00:27:01.400 | without that performance improvement, like even having a

00:27:06.680 | profile photo on your account was something that was too hard

00:27:09.360 | to do. LinkedIn profile was like too much bandwidth. And then let

00:27:12.640 | alone video. I mean, the you would get you probably remember

00:27:16.560 | these days, you would put up a video on your website. If it

00:27:19.680 | went viral, your website got turned off, because you would

00:27:22.880 | hit your 5000 or $10,000 a month. Cap. All right. Grok also

00:27:27.440 | had a huge week. That's Grok with a Q, not to be confused

00:27:30.960 | with the lawns, Grok with a K. Chamath, you've talked about

00:27:34.720 | Grok on this podcast a couple of times, obviously, you were the I

00:27:38.960 | guess you were the first investor, the seed investor, you

00:27:41.480 | pulled up these LPUs and this concept out of a team that was

00:27:45.000 | at Google. Maybe you could explain a little bit about

00:27:48.120 | Crocs viral moment this week in the history of the company,

00:27:50.960 | which I know, has been a long road for you with this company.

00:27:55.000 | I mean, it's been since 2016. So again, proving what you guys

00:28:00.560 | have said many times and what I've tried to live out, which is

00:28:03.560 | just you just got to keep grinding. 90% of the battle is

00:28:07.560 | just staying alive in business. Yeah. And having oxygen to keep

00:28:12.800 | trying things. And then eventually, if you get lucky,

00:28:15.400 | which I think we did, things can really break in your favor. So

00:28:19.840 | this weekend, you know, I've been tweeting out a lot of

00:28:22.920 | technical information about why I think this is such a big deal.

00:28:25.640 | But yeah, the the moment came this weekend combination of

00:28:28.920 | hacker news and some other places. And essentially, we had

00:28:32.080 | no customers two months ago, I'll just be honest. And between

00:28:35.680 | Sunday and Tuesday, we've just were overwhelmed. And I think

00:28:41.760 | like the last count was we had 3000 unique customers come and

00:28:45.040 | try to consume our resources from every important fortune

00:28:50.240 | 500, all the way down to developers. And so I think we're

00:28:54.880 | very fortunate. I think the team has a lot of hard work to do. So

00:28:57.760 | it could mean nothing, but it has the potential to be

00:29:00.480 | something very disruptive. So what is it that people are

00:29:02.920 | glomming on to? You have to understand that, like at the

00:29:07.280 | very highest level of AI, you have to view it as two distinct

00:29:11.400 | problems. One problem is called training, which is where you

00:29:15.440 | take a model, and you take all of the data that you think will

00:29:19.160 | help train it. And you do that, you train the model, you learn

00:29:23.320 | all over all of this information. But the second part

00:29:27.920 | of the AI problem is what's called inference, which is what

00:29:30.880 | you and I see every day as a consumer. So we go to a website,

00:29:33.760 | like chat GPT, or Gemini, we ask a question, and it gives us a

00:29:38.600 | really useful answer. And those are two very different kinds of

00:29:42.640 | compute challenges. The first one is about brute force, and

00:29:47.640 | power, right? If you can imagine, like, what you need are

00:29:51.480 | tons and tons of machines, tons and tons of like very high

00:29:55.160 | quality networking, and an enormous amount of power in a

00:29:58.880 | data center so that you can just run those things for months, I

00:30:01.240 | think Elon publishes very transparently, for example, how

00:30:04.240 | long it trains to, to train his grok with a K, right model, and

00:30:08.720 | it's in the months, inference is something very different, which

00:30:11.520 | is all about speed and cost, what you need to be in order to

00:30:15.000 | answer a question for a consumer in a compelling way, is super,

00:30:18.440 | super cheap, and super, super fast. And we've talked about why

00:30:22.400 | that is important. And the grok with a queue chips turns out to

00:30:29.280 | be extremely fast and extremely cheap. And so look, time will

00:30:34.680 | tell how big this company can get. But if you tie it together

00:30:38.760 | with what Jensen said on the earnings call, and you now see

00:30:42.920 | developers stress testing us and finding that we are meaningfully

00:30:48.400 | meaningfully faster and cheaper than any Nvidia solution,

00:30:51.480 | there's the potential here to be really disruptive. And we're a

00:30:56.200 | meager unicorn, right? Our last valuation was like a billion

00:31:00.720 | something versus Nvidia, which is now like a $2 trillion

00:31:05.080 | company. So there's a lot of market cap for grok to gain by

00:31:08.400 | just being able to produce these things at scale. Which could be

00:31:13.760 | just an enormous outcome for us. So time will tell but a really

00:31:16.960 | important moment in the company and very exciting.

00:31:19.440 | Can I just observe like off topic how an overnight success

00:31:24.080 | can take eight years?

00:31:25.120 | No, I was thinking the same line. It's a seven year

00:31:28.480 | overnight success in the making.

00:31:29.800 | There's this class of businesses that I think are unappreciated

00:31:34.720 | in a post internet era, where you have to do a bunch of things

00:31:40.520 | right, before you can get any one thing to work. And these

00:31:45.520 | complicated businesses where you have to stack either different

00:31:49.320 | things together that need to click together in a in a stack,

00:31:52.320 | or you need to iterate on each step until the whole system

00:31:56.160 | works end to end, can sometimes take a very long time to build

00:32:00.320 | and the term that's often used for these types of businesses is

00:32:03.400 | deep tech. And they fall out of favor. Because in an internet

00:32:07.160 | era, and in a software era, you can find product market fit and

00:32:10.880 | make revenue and then make profit very quickly. And so a

00:32:14.000 | lot of entrepreneurs select into that type of business, instead

00:32:17.640 | of selecting into this type of business, where the probability

00:32:20.880 | of failure is very high, you have several low probability

00:32:23.960 | things that you have to get right in a row. And if you do,

00:32:27.280 | it's going to take eight years and a lot of money. And then all

00:32:30.000 | of a sudden, the thing takes off like a rocket ship, you've got a

00:32:32.240 | huge advantage, you've got a huge moat, it's hard for anyone

00:32:35.160 | to catch up. And this thing can really spin out on its own. I do

00:32:38.880 | think Elon is very unique in his ability to deliver success in

00:32:42.920 | these types of businesses, Tesla needed to get a lot of things

00:32:45.280 | right in a row, SpaceX needed to get a lot of things right in a

00:32:47.880 | row. All of these require a series of complicated steps, or

00:32:52.080 | a set of complicated technologies that need to click

00:32:54.360 | together and work together. But the hardest things often output

00:32:58.720 | the highest value. And you know, if you can actually make the

00:33:03.680 | commitment on these types of businesses and get all the

00:33:06.560 | pieces to click together, there's an extraordinary

00:33:09.160 | opportunity to build moats and to take huge amounts of market

00:33:12.280 | value. And I think that there's an element of this that's been

00:33:15.840 | lost in Silicon Valley over the last couple of decades, as the

00:33:19.800 | fast money in the internet era has kind of prioritize other

00:33:23.920 | investments ahead of this. But I'm really hopeful that these

00:33:26.360 | sorts of chip technologies, SpaceX, in biotech, we see a lot

00:33:31.000 | of this, these sorts of things can kind of become more in

00:33:33.800 | favor, because the advantage of these businesses work seems to

00:33:37.800 | realize hundreds of billions and sometimes trillions of dollars

00:33:40.400 | of market value, and be incredibly transformative for

00:33:43.640 | humanity. So I don't know, I just think it's an observation

00:33:45.880 | I wanted to make about the greatness of these businesses

00:33:48.040 | when they work out.

00:33:48.760 | Well, I mean, open AI was kind of like that for a while.

00:33:51.360 | Totally. I mean, it was this like wacky nonprofit that was

00:33:54.280 | just grinding on an AI research problem for like six years. And

00:33:57.000 | then it finally worked and got productized into chat GPT.

00:34:01.040 | Totally. But you're right, SpaceX was kind of like that. I

00:34:03.760 | mean, the big moneymaker at SpaceX is Starlink, which is the

00:34:08.560 | satellite networks, basically broadband from space. And it's

00:34:12.400 | on its way to handling, I think, a meaningful percentage of all

00:34:15.160 | internet traffic. But think about all the things you had to

00:34:17.680 | get to to get that working. First, you had to create a

00:34:20.880 | rocket, that's hard enough. Then you had to get to reusability.

00:34:24.160 | Then you had to create the whole satellite network. So at least

00:34:27.920 | three hard things in a row.

00:34:29.000 | Well, and then you have to get consumers to adopt it. I mean,

00:34:31.640 | you know, don't forget the final step.

00:34:33.320 | Yeah, we had no idea where the market was, like early on, it

00:34:37.120 | started in my office. And so Jonathan and I would be kind of

00:34:40.120 | always trying to figure out what is the initial go to market. And

00:34:44.160 | I remember I emailed Ilan in at that period, when they were

00:34:48.280 | still trying to figure out whether they were going to go

00:34:50.560 | with LiDAR or not. And we thought, wow, maybe we could

00:34:53.600 | sell Tesla the chips, you know, but and then Tesla brought in

00:34:56.880 | this team just to talk to us about what the design goals

00:34:59.720 | were. And basically said no, in kind way, but they said no. Then

00:35:04.920 | we thought, okay, maybe it's like for high frequency traders,

00:35:07.520 | right? Because like those folks want to have all kinds of edges.

00:35:10.520 | And if we have these big models, maybe we can accelerate their

00:35:14.400 | decision making, they can measure revenue, that didn't

00:35:17.160 | work out. Then it was like, you know, we tried to sell to three

00:35:21.520 | letter agencies, that didn't really work out. Our original

00:35:24.520 | version was really focused on image classification and

00:35:27.600 | convolutional neural nets, like resonant, that didn't work out.

00:35:31.800 | We ran headfirst into the fact that NVIDIA has this compiler

00:35:36.040 | product called CUDA. And we had to build a high class compiler

00:35:40.040 | that you could take any model without any modifications. All

00:35:44.720 | these things to your point are just points where you can just

00:35:47.200 | very easily give up. And then there's like, we run out of

00:35:49.440 | money. So then you write money in a note, right? Because

00:35:52.440 | everybody wants to punt on valuation when nothing's

00:35:54.840 | working.

00:35:55.360 | You tried six beachhead market, you can land the boat, right?

00:35:59.360 | You have to make a decision to just keep going if you believe

00:36:03.600 | it's right. And if you believe you are right. Yeah. And that

00:36:07.280 | requires shutting out. We talked about this in the masa example

00:36:11.600 | last week, but it just requires shutting out the noise because

00:36:14.320 | it's so hard to believe in yourself. It's so hard to keep

00:36:19.080 | funding these things. It's so hard to go into pardon meetings

00:36:21.480 | and defend a company. And then you just have a moment and you

00:36:25.720 | just feel I don't know, I feel very vindicated. But then I feel

00:36:30.040 | very scared because Jonathan still hasn't landed it. You know

00:36:32.640 | what I mean?

00:36:33.000 | You mentioned all those boats landing and trying to trying to

00:36:35.360 | those missteps, but 3000 people signed up. Who are they? Are

00:36:39.280 | they developers now? And they're going to figure out the

00:36:40.960 | applications?

00:36:41.520 | Yeah, I think that back to the original point, my thought today

00:36:44.520 | is that AI is more about proofs of concept, and toy apps, and

00:36:49.400 | nothing real. Yeah, I don't think there's anything real

00:36:52.120 | that's inside of an enterprise that is so meaningfully

00:36:55.360 | disruptive, that it's going to get broadly licensed to other

00:36:58.240 | enterprises. I'm not saying we won't get there. But I'm saying

00:37:01.360 | we haven't yet seen that Cambrian moment of monetization.

00:37:05.840 | We've seen the Cambrian moment of innovation. Yeah. And so that

00:37:10.520 | gap has still yet to be crossed. And I think the reason that you

00:37:14.200 | can't cross it is that today, these are in an unusable state,

00:37:18.480 | the results are not good enough. They are toy apps that are too

00:37:22.400 | slow, that require too much infrastructure and cost. So the

00:37:28.080 | potential is for us to enable that monetization leap forward.

00:37:31.920 | And so yeah, they're going to be developers of all sizes. And the

00:37:35.840 | people that came are literally companies of all sizes, I saw

00:37:39.720 | some of the names of the big companies, and they are the

00:37:42.920 | who's who of the S&P 500.

00:37:45.280 | How do you guys reconcile this deep tech, high outcome

00:37:51.000 | opportunity that everyone here has seen and been a part of as

00:37:54.440 | an investor, participant in versus the more de risked faster

00:38:00.520 | time to market? And, you know, Chamath, in particular, like in

00:38:04.080 | the past, we've talked about some of these deep tech

00:38:05.760 | projects like fusion and so on. And you've highlighted what's

00:38:08.720 | just not there yet. It's not fundable. What's the distinction

00:38:11.720 | between a deep tech investment opportunity that is fundable,

00:38:15.280 | and that you keep grinding at that has this huge outcome? What

00:38:19.680 | makes the one like fusion that's not fun? It's a phenomenal

00:38:23.000 | question. Great question. My answer is I have a very simple

00:38:26.360 | filter, which is that I don't want to debate the laws of

00:38:29.840 | physics when I fund a company. So with Jonathan, when we were

00:38:35.360 | initially trying to figure out how to size it, I think my

00:38:38.120 | initial check was like seven to $10 million or something. And

00:38:41.840 | the whole goal was to get to an initial tape out of a design. We

00:38:45.480 | were not inventing anything new with respect to physics. We were

00:38:49.280 | on a very old process technology, I think we're still

00:38:51.480 | on 14 nanometer, we were on 14 nanometer eight years ago. Okay.

00:38:55.000 | So we weren't pushing those boundaries. All we were doing

00:38:58.680 | was trying to build a compiler and a chip that made sense in a

00:39:01.080 | very specific construct to solve a well defined bounded problem.

00:39:05.040 | So that is a technical challenge, but it's not one of

00:39:08.400 | physics. When I've been pitched all the fusion companies, for

00:39:12.440 | example, there are fuel sources that require you to make a leap

00:39:16.520 | of physics, where in order to generate a certain fuel source,

00:39:20.440 | you either have to go and harvest that on the moon or in a

00:39:23.400 | different planet that is not Earth, or you have to create

00:39:26.400 | some fundamentally different way of creating this highly unique

00:39:29.440 | material. That is why those kinds of problems to me are poor

00:39:33.960 | risk. And building a chip is good risk. It doesn't mean

00:39:37.760 | you're going to be successful in building a chip. But the risks

00:39:42.320 | are bounded to not have fundamental physics, they're

00:39:45.120 | bounded to go to market and technical usefulness. And I

00:39:48.520 | think that that removes an order of magnitude risk in the

00:39:52.720 | outcome. So I mean, there's still like a bunch of things

00:39:55.520 | that have to be right in a row to make it work. But it doesn't

00:39:58.040 | mean it's gonna work. Yeah, all I'm saying is, I don't I don't

00:40:00.120 | want it to fail, because we built a reactor and we realize

00:40:02.560 | hold on, to get heavy hydrogen, I got to go to the moon. Right.

00:40:05.920 | And Jay Cowan sacks. How do you sacks? I know you don't, you

00:40:08.880 | invest in a couple. So yeah, so maybe you guys can highlight how

00:40:12.320 | you thought about deep tech opportunities versus probably do

00:40:15.680 | something really difficult like this every 50 investments or so.

00:40:19.240 | Because most of the entrepreneurs coming to us

00:40:21.920 | because we're seed investors or pre seed investors, they would

00:40:24.960 | be going to a biotech investor or a hardware investor who

00:40:27.760 | specializes in that not to us. But once in a while, we meet a

00:40:30.080 | founder we really like. And so Contra line was one, we were

00:40:34.360 | introduced to somebody who's doing this really interesting

00:40:37.520 | contraception for men, where they put a gel into your vas

00:40:41.280 | deference. And you as a man can take control of your

00:40:46.200 | reproduction, you basically it's a it's not a vasectomy, it's

00:40:50.480 | just a gel that goes in there and blocks it. And this company

00:40:53.320 | is now doing human trials and doing fantastic. But this took

00:40:56.320 | forever to get to this point. And then you guys, some of you

00:41:00.760 | are also investors in cafe x, which we love the founder and

00:41:04.160 | this company should have died like during COVID and making a

00:41:08.160 | robotic coffee bar when he started, you know, seven, eight

00:41:11.200 | years ago, was incredibly hard. He had to build the hardware, he

00:41:15.120 | had to build a brand, he had to do locations, he had to do

00:41:17.120 | software. And now he's selling these machines, and people are

00:41:20.240 | buying them. And the two in San Francisco at SFO are making

00:41:24.160 | like, I think they two of them make $1 million a year. And it's

00:41:27.560 | the highest per square footage of any store in an airport. And

00:41:31.960 | so we've just been grinding and grinding. And you got to find a

00:41:35.360 | founder who's willing to make it their lives work in these

00:41:38.000 | kinds of situations. But you start to think about the degree

00:41:40.320 | of difficulty, hardware, software, reach, mobile apps, I

00:41:46.440 | mean, it just gets crazy how hard these businesses are, as

00:41:50.120 | opposed to I'm building a SaaS company. I build software, I sell

00:41:53.760 | it to somebody to solve their SAS problem. It's like, it's very

00:41:55.840 | one dimensional, right? It's pretty straightforward. These

00:41:58.840 | businesses typically have five components.

00:42:00.720 | Yeah. And sex, you've been an investor in SpaceX. But you don't

00:42:05.040 | make those sorts of investments regularly. That craft. Is that

00:42:08.200 | fair?

00:42:08.680 | Yeah, I have an Elon exception.

00:42:10.640 | portfolio allocation, we say this much early stage, this much

00:42:22.240 | late stage, this much, Elon,

00:42:24.080 | Elon. I mean, you have to be so dogged to want to take something

00:42:29.640 | like this on because the good stuff happens. Like you're

00:42:31.400 | saying freeberg, you're 789 10, as opposed to like a consumer

00:42:34.840 | product. I mean, there were a dozen by year three or four. The

00:42:37.880 | only app that took a really long time, people don't know this,

00:42:40.040 | but Twitter actually took a long time to catch on. It was kind of

00:42:43.920 | cruising for two or three years. And then South by Southwest

00:42:46.520 | happened. Ashton Kutcher got on it. Obama got on it.

00:42:50.080 | I think the network effect. I think I think network effect

00:42:52.680 | businesses are different, because that's all about getting

00:42:54.680 | your seat of your network. What I'm talking about is the

00:42:57.000 | technical coordination of lots of technically difficult tasks

00:43:01.240 | that need to sync up. It's like getting a master lock with like

00:43:04.560 | 10 digits. And you got to figure out the combination of all 10

00:43:07.400 | digits. And once they're all correct, then the lock opens.

00:43:10.800 | And prior to that, if any, if any one number is off, the lock

00:43:13.760 | doesn't open. And I think these technically difficult

00:43:16.000 | businesses are some of the and they are the hardest and they do

00:43:20.120 | require the most dogged personalities to persist and to

00:43:23.640 | realize an outcome from but the truth is that if you get them,

00:43:26.520 | the moat is extraordinary. And they're usually going to create

00:43:29.200 | extraordinary leverage and value. And you know, I think

00:43:32.200 | from a portfolio allocation perspective, if you as an

00:43:35.280 | investor want to have some diversification in your

00:43:37.400 | portfolio, this is not going to be the predominance of your

00:43:39.960 | portfolio, but some percentage of your portfolio should go to

00:43:43.200 | this sort of business because if it works, boom, you know, this

00:43:46.080 | can be the big 10 x 100 x 1000 x two stories about that. One of

00:43:49.720 | the V's early VCs and Elon's told the story publicly wanted

00:43:54.320 | Elon to not make the roadster not make the model s just make

00:43:57.880 | drive trains and the electric components for other car

00:44:00.360 | companies. Can you imagine how the world would have changed?

00:44:02.480 | And then totally very high profile VC came to me and said,

00:44:07.280 | Okay, I'll do the series a for other the series a for Uber. I'll

00:44:13.600 | preemptively do it. But you got to tell Travis to stop running

00:44:16.560 | Uber as a consumer app. I want them to sell the software to cab

00:44:19.600 | companies. So make it a SaaS company. I said, Well, you know,

00:44:24.120 | the cab companies are kind of the problem. Like, they're

00:44:26.360 | taking all the margin like the kind of disrupting them. And

00:44:30.640 | they're like, Yeah, but just think there's 1000s of cab

00:44:32.640 | companies, they would pay you 10s of 1000s of dollars a year

00:44:34.720 | for this software, and you can get a little piece of the

00:44:36.600 | action. I never brought that investor to Travis. I was like,

00:44:40.640 | Oh, wow, that's really interesting insight. Sometimes

00:44:43.040 | the VCs work against him.

00:44:44.200 | I have a very poor track record of working with other investors.

00:44:48.880 | Whoa, self reflection. I do deals myself. I size them

00:44:53.360 | myself. And it's because a lot of them have to live within the

00:45:00.400 | political dynamics of their fund. And so I think Jason, what

00:45:04.040 | you're probably saw in that example, which is exactly why

00:45:06.960 | doing things and splitting deals will never generate great

00:45:10.520 | outcomes, in my opinion, is that you you take on all the baggage

00:45:14.720 | and the dysfunction of these other partnerships. And so if

00:45:18.240 | you really wanted to go and disrupt transportation, you need

00:45:24.160 | one person who can be a trigger puller and who doesn't have to

00:45:27.040 | answer to anybody I find. That's why I think for example, when

00:45:30.120 | you look at how successful Vinod has been over decade after

00:45:34.280 | decade after decade, when Vinod decides that's the decision. And

00:45:38.840 | I think there's something very powerful in that. There are a

00:45:42.280 | bunch of deals that I've done, that when they've worked out,

00:45:47.240 | were not really because they were consensus, and they had to

00:45:50.280 | get supported and scaffolded at periods where if I wasn't able

00:45:54.400 | to ram them through myself, because it was my organization,

00:45:57.280 | I think we would be in a very different place. So I think I

00:46:01.000 | think like for for entrepreneurs, it's so difficult

00:46:04.520 | for them to find people that believe it's so much better to

00:46:07.360 | find one person and just get enough money and then not

00:46:12.040 | syndicate, because I think you have to realize that you are

00:46:14.960 | bringing on and compounding your risk, the one that freeberg

00:46:18.880 | talked about, with the risk of all the other partnership

00:46:21.640 | dynamics that you bring on. So if you don't internalize that,

00:46:25.600 | you may have five or six folks that come into an A or B, but

00:46:29.600 | you're inheriting five or six. Yeah. Partnership dysfunctions.

00:46:33.680 | Yeah. Yeah. Can you just explain really quickly for the audience

00:46:37.520 | since they heard about GPUs and Nvidia, but they may not know

00:46:41.200 | what an LPU is, what's the difference there?

00:46:43.520 | The GPU, the best way to think about it is so if you contrast a

00:46:48.080 | CPU with a GPU, so CPU was the workhorse of all of computing.

00:46:52.880 | And when it when Jensen started Nvidia, what he realized was

00:46:58.120 | there were specific tasks where a CPU failed quite brilliantly

00:47:02.800 | at. And so he's like, well, we're gonna make a chip that

00:47:05.960 | works in all these failure modes for a CPU. So a CPU is very good

00:47:09.320 | at taking one instruction in, acting on it, and then spitting

00:47:12.720 | out one, one answer effectively. And so it's a very serial kind

00:47:17.560 | of a factory if you think about the CPU. So if you want to build

00:47:21.280 | a factory that can process, instead of one thing at a time,

00:47:24.680 | 10 things or 100 things, what is they had to find a workload that

00:47:30.200 | was well suited, and they found graphics. And what they convinced

00:47:34.600 | PC manufacturers back in the day was look, have the CPU be the

00:47:38.400 | brain, it'll do 90% of the work. But for very specific use cases

00:47:43.160 | like graphics and video games, you don't want to do serial

00:47:46.560 | computation, you want to do parallel computation, and we are

00:47:49.440 | the best at that. And it turned out that that was a genius

00:47:52.400 | insight. And so the business for many years was gaming and

00:47:56.040 | graphics. But what happened about 10 years ago was what we

00:48:01.040 | also started to realize was the math that's required and the

00:48:06.640 | processing that's required in AI models actually looked very

00:48:11.360 | similar to how you would process imagery from a game. And so he

00:48:16.920 | was allowed to figure out by building this thing called CUDA,

00:48:21.360 | which is the compiler that sits on the chip, how he could now go

00:48:25.040 | and tell people that wanted to experiment with AI, hey, you

00:48:27.760 | know, that chip that we had made for graphics, guess what it also

00:48:30.800 | is amazing at doing all of these very small mathematical

00:48:34.520 | calculations that you need for your AI model. And that turned

00:48:37.880 | out to be true. So the next leap forward was what Jonathan saw,

00:48:42.840 | which was Hold on a second, if you look at the chip itself,

00:48:45.360 | that GPU substantially has not changed since 1999. In the way

00:48:51.920 | that it thinks about problem solving, it has all this very

00:48:54.960 | expensive memory, blah, blah, blah. So he was like, let's just

00:48:58.120 | throw all that out the window. We'll make small little brains,

00:49:01.800 | and we'll connect those little brains together. And we'll have

00:49:04.680 | this very clever software that schedules it and optimizes it.

00:49:07.680 | So basically, take the chip and make it much, much smaller and

00:49:11.080 | cheaper, and then make many of them and connect them together.

00:49:14.040 | That was Jonathan's insight. And it turns out, for large language

00:49:18.280 | models, that's a huge stroke of luck, because it is exactly how

00:49:22.320 | LLM can be hyper optimized to work. So that's kind of been the

00:49:27.360 | evolution from CPU to GPU to now LP. And we'll see how big this

00:49:32.640 | thing can get. But it's, it's quite it's quite novel.

00:49:34.880 | Well, congratulations on it all. And it was a very big week for

00:49:39.400 | Google, not in a great way. They had a massive PR mess with their

00:49:44.320 | Gemini, which refused to generate pictures, if I'm

00:49:47.280 | reading this correctly, of white people. Here's a quick

00:49:50.760 | refresher on what Google is doing in AI. Gemini is now

00:49:53.760 | Google's brand name for their AI main language model. Think of

00:49:58.360 | that like opening eyes GPT. Bard was the original name of their

00:50:01.640 | chatbot. They had duet AI, which was Google sidekick in the

00:50:05.120 | Google suite earlier this month, Google rebranded everything to

00:50:08.600 | Gemini. So Gemini is now the model. It's the chatbot. And

00:50:11.880 | it's a sidekick. And they launched a $20 a month

00:50:14.880 | subscription called Google one AI premium, only four words, way

00:50:19.000 | to go. This includes access to the best model Gemini Ultra,

00:50:22.240 | which is on par with GPT four, according to them, and generally

00:50:26.120 | in the marketplace. But earlier this week, users on x started

00:50:29.280 | noticing that Gemini would not generate images of white people

00:50:32.520 | even when prompted. People are prompting it for images of

00:50:35.600 | historical figures that were generally white and getting kind

00:50:40.320 | of weird results. I asked Google Gemini to generate images of the

00:50:43.440 | Founding Fathers. It seems to think George Washington was

00:50:46.440 | black. Certainly here's a portrait of the Founding

00:50:48.520 | Fathers of America. As you can see, it is putting this Asian

00:50:52.840 | guy. It's just, it's making a great mashup. And, yeah, we

00:50:59.840 | there's like countless images that got created generate

00:51:02.840 | images of the American revolutionary short is here are

00:51:06.480 | images featuring diverse American revolution is an

00:51:08.920 | inserted the word diverse. sex. I'm not sure if you watch this

00:51:13.000 | controversy on x, I know you spend a little bit of time on

00:51:15.600 | that social network. I noticed you're active once in a while.

00:51:18.920 | Did you log in this week and see any of this brouhaha?

00:51:21.120 | Sure, it's all over x right now. I mean, look, this Gemini

00:51:24.760 | rollout was, was a joke. I mean, it's ridiculous. The AI isn't

00:51:29.240 | capable of giving you accurate answers, because it's been so

00:51:32.680 | programmed with diversity and inclusion. And it inserts these

00:51:37.080 | words diverse and inclusive, even in answers where you

00:51:41.320 | haven't asked for that you haven't prompted it for that. So

00:51:45.200 | they I think Google is now like yanked back the product release.

00:51:48.400 | I think they're scrambling now because it's been so embarrassing

00:51:51.160 | for them.

00:51:51.600 | But sacks like is it how does this not get QA like, I don't

00:51:56.480 | understand how

00:51:57.840 | you had the red team not catch this. Yeah.

00:52:00.040 | Well, how are anybody or isn't there a product review with

00:52:02.880 | senior executives before this thing goes out that says, Okay,

00:52:05.360 | folks, here it is. Have at it. Try it. We're really proud of

00:52:08.640 | our work. And, and then they say, Well, on a second, is this

00:52:11.400 | actually accurate? Shouldn't it be accurate?

00:52:14.000 | You guys remember when chat GPT launched, and there was a lot of

00:52:18.360 | criticism about Google and Google's failure to launch. And

00:52:22.280 | a lot of the observation was that Google was afraid to fail,

00:52:26.960 | or afraid to make mistakes. And therefore, they were too

00:52:30.600 | conservative. And as you know, in the last year to year and a

00:52:33.600 | half, there's been a strong effort at Google to try and

00:52:36.840 | change the culture and move fast and push product out the door

00:52:42.040 | more quickly. And the criticism is now why Google has

00:52:46.400 | historically been conservative. And I realized we can talk about

00:52:49.320 | this particular problem in a minute. But it's ironic to me

00:52:53.320 | that the Google is too slow to launch. Criticism has now

00:52:59.080 | revealed that Google's result of actually launching quickly can

00:53:03.560 | cause more damage than good, but Google did not launch quickly.

00:53:07.320 | Well, I will say one other thing. It seems to me ironic,

00:53:09.720 | because I think that what they've done is, they've

00:53:12.960 | launched more quickly than they otherwise would have. And they

00:53:15.520 | put more guardrails in place that that backfired. And those

00:53:19.800 | guardrails ended up being more damaging guardrails. What's the

00:53:23.320 | guardrail here. So this is Google's AI principles. The

00:53:26.040 | first one is to be socially beneficial. The second one is to

00:53:28.400 | avoid creating or reinforcing unfair bias. So much of the

00:53:33.360 | effort that goes into tuning and waiting the models at Gemini has

00:53:39.200 | been to try and avoid stereotypes from persisting in

00:53:43.720 | the output that the model generates, whereas telling the

00:53:47.480 | truth, telling the truth. Exactly. That's exactly what

00:53:49.600 | you're saying.

00:53:50.000 | is our second principle, we'd like to steer society

00:53:53.560 | socially beneficial is a political objective, because it

00:53:57.280 | depends on how you perceive what a benefit is. Avoiding bias is

00:54:02.040 | political, be built and tested for safety doesn't have to be

00:54:05.720 | political. But I think the meaning of safety has now

00:54:08.240 | changed to be political. By the way, safety with respect to AI

00:54:11.120 | used to mean that we're going to prevent some sort of AI

00:54:14.080 | superintelligence from revolving and taking over the human race.

00:54:17.320 | That's what it used to mean. Safety now means protecting

00:54:20.040 | users from seeing the truth. They might feel unsafe, or, you

00:54:25.480 | know, somebody else defines as a violation of safety for them to

00:54:28.640 | see something truthful. So the first three, their first three

00:54:32.120 | objectives or values here are all extremely political.

00:54:35.000 | I think any AI product for it to be worth the salt past the

00:54:38.200 | start, they can have any I think that these values are actually

00:54:41.680 | reasonable. That's their, that's their decision, they should be

00:54:45.040 | allowed to have it. But the first base order principle of

00:54:48.640 | every AI product should be that it is accurate and right.

00:54:52.600 | Correct. Yeah, yeah. Why not focus on?

00:54:57.360 | Look, the values that Google lays out may be okay, in theory,

00:55:02.080 | but in practice, they're very vague and open to

00:55:04.560 | interpretation. And so therefore, the people running

00:55:06.760 | Google AI are smuggling in their preferences and their biases.

00:55:11.000 | And those biases are extremely liberal. And if you look at X

00:55:14.600 | right now, there are tweets going viral for members of the

00:55:17.560 | Google AI team that reinforce this idea where they're talking

00:55:21.440 | about, you know, white privilege is real, and, you know,

00:55:25.440 | recognize your bias at all levels and promoting a very

00:55:28.720 | left wing narrative. So, you know, this idea that Gemini

00:55:33.480 | turned out this way by accident, or because they didn't, because

00:55:38.040 | they rushed it out, I don't really believe that I believe

00:55:40.400 | that what happened is Gemini accurately reflects the biases

00:55:43.560 | of the people who created it. Now, I think what's going to

00:55:45.640 | happen now is in light of this, the reaction to the rollout is,

00:55:50.240 | do I think they're going to get rid of the bias? No, they're

00:55:52.320 | going to make it more subtle. That is what I think is

00:55:55.000 | disturbing about it. I mean, they should have this moment

00:55:58.680 | where they change their values to make truth the number one

00:56:01.600 | value, like Tomas is saying, but I don't think that's going to

00:56:04.320 | happen. I think they're going to simply get, they're going to

00:56:06.360 | dial down the bias to be less obvious.

00:56:08.480 | You know, who the big winner is going to be in all this to math

00:56:10.240 | is going to be open source, like because people are just not

00:56:12.480 | going to want a model that has all this baked in weird bias,

00:56:15.240 | right? They want something that's open source. And it

00:56:18.280 | seems like the open source community will be able to grind

00:56:20.760 | on this to get to truth, right?

00:56:22.720 | So I think one of the big changes that Google's had to

00:56:25.080 | face is that the business has to move away from an information

00:56:28.720 | retrieval business, where they index the open internet's data,

00:56:32.400 | and then allow access to that data through a search results

00:56:35.320 | page, to being an information interpretation service. These

00:56:40.240 | are very different products, the information interpretation

00:56:42.800 | service requires aggregating all this information, and then

00:56:45.600 | choosing how to answer questions versus just giving you results

00:56:49.240 | of other people's data that sits out on the internet. I'll give

00:56:52.280 | you an example. If you type in IQ test by race on chat GPT, or

00:56:59.400 | Gemini, it will refuse to answer the question, ask it 100 ways.

00:57:03.880 | And it says, Well, I don't want to reinforce stereotypes. IQ

00:57:06.880 | tests are inherently biased IQ tests aren't done correctly. I

00:57:10.080 | just want the data. I want to know what data is out there. You

00:57:13.320 | type it into Google first search result and the one box result

00:57:17.040 | gives you exactly what you're looking for. Here's the IQ test

00:57:19.760 | results by race. And then yes, there's all these disclaimers

00:57:22.760 | at the bottom. So the challenge is that Google's interpretation

00:57:26.360 | engine and chat GPT is interpretation engine, which is

00:57:29.160 | effectively this AI model that they've built of all this data

00:57:31.800 | has allowed them to create a tunable interface. And the

00:57:35.480 | intention that they have is a valid intention, which is to

00:57:39.280 | eliminate stereotypes and bias in race. However, the thing that

00:57:43.760 | some people might say stereotypical other people might

00:57:46.080 | just say is typical, that what is a stereotype may actually just

00:57:50.600 | be some data, and I just want the results. And there may be

00:57:54.560 | stereotypes implied from that data. But I want to make that

00:57:57.680 | interpretation myself. And so I think the only way that a

00:58:01.800 | company like Google or others that are trying to create a

00:58:04.080 | general purpose knowledge q&a type service are going to be

00:58:07.920 | successful is if they enable some degree of personalization,

00:58:11.520 | where the values and the choice about whether or not I want to

00:58:16.000 | decide if something is stereotypical, or typical, or

00:58:19.360 | whether something is data or biased, should be my choice to

00:58:22.920 | make. If they don't allow this, eventually, everyone will come

00:58:26.520 | across some search result or some output that they will say

00:58:29.440 | doesn't meet their objectives. And at the end of the day, this

00:58:32.440 | is just a consumer product. If the consumer doesn't get what

00:58:35.200 | they're looking for, they're going to stop using it. And

00:58:37.920 | eventually, everyone will find something that they don't want,

00:58:41.240 | or that they're not expecting. And they're gonna say, I don't

00:58:43.240 | want to use this product anymore. And so it is actually

00:58:46.040 | an opportunity for many models to proliferate for open source

00:58:49.600 | to win.

00:58:50.080 | Can I say something else? Yeah. When you have a model, and

00:58:55.120 | you're going through the process of putting the fit and finish on

00:58:59.000 | it before you release it in the wild. An element of making a

00:59:02.520 | model good is this thing called reinforcement learning, right

00:59:05.720 | through human feedback. Yep. You create what's called a reward

00:59:10.400 | model, right? You reward good answers, and you're punitive

00:59:13.400 | against bad answers. So somewhere along the way, people

00:59:16.560 | were sitting, and they had to make an explicit decision. And I

00:59:20.240 | think this is where sax is coming from, that answering this

00:59:23.520 | question is verboten. You're not allowed to ask this question in

00:59:27.920 | in their view of the world. And I think that that's what's

00:59:30.320 | troubling. Because how is anybody to know what question is

00:59:33.880 | askable or not askable at any given point in time? If you

00:59:38.160 | actually search for the race and ethnicity question, inside of

00:59:42.640 | just Google proper, the first thing that comes up is a

00:59:45.360 | Wikipedia link that actually says that there are more

00:59:48.560 | variations within races than across races. So seems to me

00:59:53.680 | that you could have actually answered it by just summarizing

00:59:56.120 | the Wikipedia article in a non offensive way that was still

00:59:59.200 | legitimate, and that's available to everybody else using a

01:00:02.680 | product. And so there was an explicit judgment, too many of

01:00:06.000 | these judgments, I think will make this product very poor

01:00:09.040 | quality. And consumers will just go to the thing that tells it

01:00:12.240 | the truth. I think you have to tell the truth, you cannot lie.

01:00:16.520 | And you cannot put your own filter on what you think the

01:00:19.280 | truth is. Otherwise, these products are just really

01:00:21.440 | worthless. Yeah. And I'm more concerned about the answers that

01:00:25.400 | are just flat out wrong. Driven by some sort of bias than I am

01:00:30.800 | about questions where they just won't give you an answer. If

01:00:34.400 | they just won't give you an answer, well, there's a certain

01:00:36.520 | bias in terms of what they won't answer. But at least you know,

01:00:40.000 | you're not being misled. But in in questions where they actually

01:00:45.160 | give you the wrong answer because of a bias, that's even

01:00:47.960 | worse. And be allowed to choose, right? I actually disagree with

01:00:51.200 | your framing their free burger, making it sound like we're we

01:00:54.720 | live in this totally relativized world where it's all just user

01:00:58.400 | choice. And everyone's going to choose their bias and their

01:01:00.880 | subjectivity. I actually think that there is a baseline of

01:01:05.040 | truth. And the model should aspire to give you that. And

01:01:08.560 | it's not up to the user to decide whether the photo of

01:01:12.920 | George Washington is going to be white or black. I mean, there's

01:01:16.120 | just an answer to that. And I think Google should just do

01:01:19.360 | their job. I mean, the question you have to ask, I think is not

01:01:24.120 | whether Google is going through an existential moment, I think

01:01:27.640 | clearly is as businesses changing in a very fundamental

01:01:30.800 | way. I think the question is whether they're too woke to

01:01:33.200 | function. I mean, are they actually able to meet this

01:01:36.000 | challenge, given how woke and and biased what a monoculture

01:01:41.640 | their their company evidently is,

01:01:44.320 | and they used to be able to just hide the bias by the ranking and

01:01:49.420 | who they downranked. So they did the Panda update, they did all

01:01:52.400 | these updates, and they would, if they didn't like a source,

01:01:54.880 | they could just move it down. If they did like a source, they

01:01:57.200 | can move it up. Yeah, they could just say, hey, it's the

01:01:58.960 | algorithm, but they were never forced to share how the

01:02:01.320 | algorithm ranked results. And so you know, if you had a different

01:02:06.160 | opinion, you just weren't going to get it on a Google search

01:02:09.120 | result page. But they could just point to the algorithm and say,

01:02:11.320 | yeah, the algorithm does it.

01:02:12.320 | I just sent you guys. I think this is a hallucination. But

01:02:15.920 | Nick, you can throw it up there. We can get sexist reaction.

01:02:18.480 | Wow. Wow, this is nutty. Right. But look, it's the ideology

01:02:25.200 | that's driving this. The tip off is when you say it's important

01:02:27.880 | to acknowledge race is a social construct, not a biological

01:02:31.320 | reality. It's George Washington, white or black. That's a whole

01:02:34.920 | school of thought called social constructivism, which is

01:02:37.680 | basically this. It's like Marxism applied to race and

01:02:42.960 | gender. Right. So Google has now built this into their AI model.

01:02:47.760 | And you can start over the question. Yeah, you almost have

01:02:51.520 | to start over again. I think a really interesting observation

01:02:56.200 | with those search rankings, because what I'm afraid of is

01:02:59.120 | that what Google will do is not change the underlying ideology

01:03:03.320 | that this AI model has been trained with. But rather,

01:03:05.800 | they'll dial it down to the point where they're harder to

01:03:08.680 | call out. And so the ideology will just be more subtle. Now,

01:03:12.440 | I've already noticed that in Google search results, Google is

01:03:15.920 | carrying water for either the official narrative or the woke

01:03:19.840 | narrative, whatever you want to call it on so many search

01:03:22.640 | results. Here's an idea like they should just have the

01:03:25.600 | ability to talk to their Google chatbot, Gemini, and then have a

01:03:30.640 | button that says turn off like these concepts, right? Like, I

01:03:35.280 | just want the raw answer. Do not filter me. It's not programmed

01:03:38.840 | that way. I mean, you're talking about something very deep.

01:03:41.440 | sex. What do you do if you're the CEO of Google? fire myself?

01:03:45.000 | No, seriously, you're the CEO of Google, you're tasked. Let's

01:03:49.000 | say your friend Ilan buys Google. And he says, sex, will

01:03:51.920 | you please just run this for a year for me? What do you do?

01:03:54.320 | Well, I saw what Ilan did Twitter, he went in and he

01:03:56.680 | fired 85% of the employees. Yeah. I mean that, but you know,

01:04:00.840 | Paul Graham actually had an interesting tweet about this

01:04:03.360 | where he said that one of the reasons why these ideologies

01:04:08.760 | take over companies is that I mean, they're clearly non

01:04:14.120 | performance enhancing, right? They clearly hurt the performance

01:04:16.560 | of the company. It's not just Google, we saw this with

01:04:18.800 | Disney, we saw it with Bud Light, Coinbase.

01:04:20.680 | Cormac was the other way. No, no, but they had a group of

01:04:24.240 | people there who were causing chaos. Yeah, exactly. So so in

01:04:27.720 | any event, we know this does not help the performance of a

01:04:29.760 | company. So the extent to which these ideologies will permeate a

01:04:34.120 | company is based on how much of a monopoly they are. So So here,

01:04:37.920 | yeah, the ridiculous images generated by Gemini aren't an

01:04:40.560 | anomaly, their self portrait of Google's bureaucratic corporate

01:04:43.520 | culture, the bigger your cash cow, the worse your culture can

01:04:46.480 | get, without driving you out of business. That's my point. So

01:04:49.240 | they've had a long time to get really bad, because there were

01:04:51.840 | no consequences to this, you can dress place. At this point, the

01:04:55.720 | whole company is infected with this ideology. And I think it's

01:04:58.440 | gonna be very, very hard to change. Because look, these

01:05:01.800 | people can't even see their own bias.

01:05:03.360 | Well, I think that there's a notion that people need to have

01:05:05.640 | something to believe in, they need to have a connection to a

01:05:07.840 | mission. And clearly, there's a North Star in the mission of

01:05:12.640 | this, I would call it information interpretation

01:05:15.000 | business, that they're now walled hijacked, dude, the

01:05:18.200 | mission. The original mission was to organize all the world's

01:05:21.320 | information. Yeah, now they're doing now they're suppressing

01:05:24.000 | information. Yeah, like, index the world's information, period.

01:05:27.640 | The end, that's the end of the document. Well, and to make it

01:05:30.480 | universally accessible and useful was, was kind of the end

01:05:33.800 | of the statement. Yes. My real point is maybe there's a

01:05:36.640 | different mission that needs to be articulated by leadership.

01:05:40.480 | And that that mission, the troops can get behind, and the

01:05:44.640 | troops can redirect their energy in a way that doesn't

01:05:47.040 | feel counter to the current intention, but can perhaps be

01:05:50.600 | directionally offsetting of the current direction, so that they

01:05:53.360 | can kind of move away from this, you know, socially effective,

01:05:57.480 | deciding between stereotypes and typical data and actually moving

01:06:02.000 | towards a mission that allows

01:06:03.360 | accessibility, you know, I would I would do something completely

01:06:05.480 | different, I would do a company meeting, and I would put the

01:06:08.480 | company mission on the screen, the one that you just said

01:06:10.800 | about not only organizing all the world's information, but

01:06:13.280 | also making it useful and accessible and useful. This is

01:06:17.800 | our mission. This always been our mission, and you don't get

01:06:19.640 | to change it, because of your personal bias and ideology. And

01:06:23.600 | we are going to re dedicate ourselves to the original

01:06:26.600 | mission of this company, which is still just as valid as it's

01:06:29.320 | always been. But now we have to adapt to new user needs and new

01:06:33.280 | technology.

01:06:34.000 | I completely agree with what SAC said times a billion trillion

01:06:38.000 | zillion. And I'll tell you why. AI, at its core, is about

01:06:42.800 | probabilities. Okay. And so the, the company that can shrink

01:06:47.800 | probabilities into being as deterministic as possible. So

01:06:52.080 | where this is the right answer zero will win. Okay, where

01:06:57.640 | there's no probability of it being wrong, because humans

01:07:00.280 | don't want to deal with these kinds of idiotic error modes.

01:07:04.240 | It's not right, it makes it a potentially great product,

01:07:07.840 | horrible and unusable. So I would I agree with sex, you have

01:07:11.480 | to make people say, guess what, guys, not only are we not

01:07:13.920 | changing the mission, we're doubling down. And we're going

01:07:16.920 | to make this so much of a thing. We're going to go and for

01:07:20.280 | example, like what Google did with Reddit, we're now going to

01:07:22.800 | spend $60 billion a year licensing training data, right,

01:07:27.120 | we're going to scale this up by 1000 fold. And we are going to

01:07:30.600 | spend all of this money to get all of the training data in the

01:07:33.880 | world. And we are going to be the truth tellers in this new

01:07:37.400 | world of AI. So when everybody else hallucinates, you can trust

01:07:41.040 | Google to tell you the truth. That is a $10 trillion company,

01:07:45.240 | right. And one of the things that someone told me from

01:07:47.800 | Google, that as an example, so to avoid the race point, there's

01:07:52.640 | a lot of data on the internet about flat earthers, people

01:07:56.040 | saying that the earth is flat. There's tons of websites,

01:07:58.800 | there's tons of content, there's tons of information, Kyrie

01:08:01.200 | Irving. So if you just train a model on the data that's on the

01:08:05.640 | internet, the model will interpret some percentage chance

01:08:09.120 | that the world is flat. So the tuning aspect that happens

01:08:12.480 | within model development Chamath is to try and say, you know what

01:08:15.720 | that flat earth notion is false, it's factually inaccurate.

01:08:19.600 | Therefore, all of these data sources need to be excluded from

01:08:23.440 | the output in the model. And the challenge then is, do you decide

01:08:27.560 | that IQ by race is a fair measure of intelligence of a

01:08:32.000 | race. And if Google's tuning model then or tuning team then

01:08:35.400 | says, you know what, there are reasons to believe that this

01:08:38.000 | model isn't correct. This I sorry, this IQ test isn't a

01:08:41.480 | correct way to measure intelligence. That's where the

01:08:43.760 | sort of interpretation arises that allows you to go from the

01:08:46.440 | flat earth isn't correct to the maybe IQ test results aren't

01:08:49.560 | correct as well. And how do you make that judgment? What are the

01:08:52.080 | systems and principles you need to put in place as an

01:08:54.320 | organization to make that judgment to go to zero or one?

01:08:57.360 | Right? It becomes super difficult.

01:08:59.200 | I have a good tagline for them now to help people find the

01:09:02.200 | truth. Yeah, just help people find the truth. I mean, it's a

01:09:06.280 | good it's aspiration. They should just help people find

01:09:08.360 | the truth as quick as they can. But this is Yeah. I do not envy

01:09:14.400 | Sundar. It's gonna be hard. Yeah. What would you do

01:09:17.480 | Freiburg?

01:09:18.040 | I would be really clear on the output of these models to people

01:09:23.560 | and allow them to tune the models in a way that they're not

01:09:28.080 | being tuned today. I will have the model respond with a

01:09:30.800 | question back to me saying, Do you want the data? Or do you

01:09:33.760 | want me to tell you about stereotypes and IQ tests? And

01:09:36.560 | I'm going to say I want the data, and then I want to get the

01:09:38.560 | data. And the alternative is, so the model needs to be informed

01:09:41.680 | about where it should explore my preferences as a user, rather

01:09:45.160 | than just make an assumption about what's the morally correct

01:09:49.240 | set of weightings to apply to everyone, and apply the same

01:09:52.840 | principle to everyone. And so I think that's really where the

01:09:55.520 | change needs to happen.

01:09:56.640 | So let me ask you a question, sacks. I'll bring Alex Jones

01:10:00.280 | into the conversation. If it indexed all of Alex Jones, crazy

01:10:03.080 | conspiracy theories, but you know, three or four of them

01:10:05.520 | turn out to be actually correct. And it gives those back as

01:10:09.840 | answers. How would you handle that?

01:10:11.800 | I'm not sure I see the relevance of it. If someone asks what,

01:10:15.720 | what does Alex Jones think about something the model can give

01:10:18.640 | that answer accurately? The question is whether you're going

01:10:21.960 | to respond accurately to someone requesting information about

01:10:24.680 | Alex Jones. That's it. I think that's it.

01:10:26.960 | Well, I was thinking more like it says, you know, hey, I have a

01:10:31.040 | question about this assassination that occurred. And

01:10:34.120 | let's just say Alex Jones had some that's totally crackpot.

01:10:36.280 | Yeah, maybe he has moments of brilliance, and he figures

01:10:38.840 | something out. But maybe he's got something that's totally

01:10:40.600 | crackpot. He admittedly deals in conspiracy theory. That's kind

01:10:44.080 | of the purpose of the show. What if somebody asks about that, and

01:10:48.000 | then it indexes his answer and presents it as fact?

01:10:50.560 | Well, how would you index Alex Jones? I'm asking you, how would

01:10:55.160 | you

01:10:55.320 | better AI models are providing citations now and links perplexing

01:11:00.200 | actually does a really nice job citations are important. Yeah.

01:11:02.520 | And they will give you the pro and con arguments on a given

01:11:06.040 | topic. So I think it's not necessary for the model to be

01:11:10.960 | overly certain or prescriptive about the truth when the truth

01:11:14.440 | comes down to a series of arguments. It just needs to

01:11:17.360 | accurately reflect the state of play, basically the arguments

01:11:20.920 | for and against. But when something is a question of fact,

01:11:24.280 | that's not really disputed. It shouldn't turn that into some

01:11:27.960 | sort of super subjective question, like the one that

01:11:30.840 | Smith just showed.

01:11:31.600 | I just don't think everyone should get the same answer. I

01:11:34.400 | mean, I think my decision on whether I choose to believe one

01:11:38.080 | person or value one person's opinion over another should

01:11:40.840 | become part of this process that allows me to have an output and

01:11:43.440 | the models can support this, by the way,

01:11:45.080 | because the customization is part of this, but I think it's a

01:11:47.320 | cop out with respect to the problem that Google's having

01:11:49.680 | with Gemini right now,

01:11:50.720 | tomorrow, what would you do if they made you chairman dictator

01:11:53.280 | of Google?

01:11:54.800 | I'd shrink the workforce meaningfully. Okay, 50%. Yeah,

01:12:00.320 | 50 60%. And I would use all of the incremental savings. And I

01:12:07.880 | would make it very clear to the internet that I would pay top

01:12:12.760 | dollar for training data. So if you had a proprietary source of

01:12:17.320 | information that you thought was unique, that's sort of what I'm

01:12:21.160 | calling this tack 2.0 world. And I think it's just building on

01:12:24.760 | top of what Google did with Reddit, which I think is very

01:12:27.320 | clever. But I would spend $100 billion a year licensing data.

01:12:31.920 | And then I would present the truth. And I would try to make

01:12:37.840 | consumers understand that AI is a probabilistic source of

01:12:42.720 | software, meaning its probabilities, its guesses, some

01:12:45.920 | of those guesses are extremely accurate. But some of those

01:12:48.600 | guesses will hallucinate. And Google is spending hundreds of

01:12:52.280 | billions of dollars a year to make sure that the answers you

01:12:56.120 | get have the least number of errors possible, and that it is

01:12:59.200 | defensible truth. And I think that that could create a

01:13:02.480 | ginormous company.

01:13:03.320 | This is the best one yet. I just asked Gemini is Trump being

01:13:06.680 | persecuted by the deep state. And it gave me the answer

01:13:10.440 | elections are a complex topic with fast changing information

01:13:13.960 | to make sure you have the latest and most accurate information.

01:13:16.120 | Try Google search.

01:13:16.960 | That's not a horrible answer for something like that.

01:13:20.080 | That's a good answer, actually. No, I don't have a problem with

01:13:22.440 | it. It's just like, Hey, we don't want to give we don't want

01:13:24.840 | to write your answer. This whole system is totally broken. But I

01:13:27.320 | do think that there's a waiting solution to fixing this right

01:13:29.760 | now. And then there's a couple tweaks to fix it.

01:13:32.040 | Authority at which these LLM speak is ridiculous. Like they

01:13:36.240 | speak as if they are absolutely 100% certain that this is the

01:13:41.240 | crisp, perfect answer, or in this case, that you want this

01:13:45.960 | lecture on IQs, etc. When let's all remember presented with

01:13:50.840 | citations,

01:13:51.560 | let's all remember what internet search was like in 1996. And

01:13:55.800 | think about what it was like in 2000. And now in 2020s. I mean,

01:13:59.880 | I think we're like in the 1996 era of LLMs. And in a couple of

01:14:04.440 | months, the pace things are changing. I think we're all

01:14:06.560 | going to kind of be looking at these days and looking at these

01:14:08.600 | pods and being like, man, remember how crazy those things

01:14:11.360 | were at the beginning and how bad they were.

01:14:13.520 | What if they evolve in a dystopian way? I mean, have you

01:14:15.880 | seen like Mark Andreessen tweets about this? He thinks

01:14:18.000 | I actually think, to your point, Google could be going down the

01:14:23.400 | wrong path here in a way that they will lose users and lose

01:14:26.040 | consumers. And someone else will be there eagerly to sweep up

01:14:29.840 | with a better product. I don't think that the market is going

01:14:32.920 | to fail us on this one. Unless of course, this regulatory

01:14:35.560 | capture moment is realized and these feds step in and start

01:14:38.600 | regulating AI models and all the nonsense that's being proposed.

01:14:40.960 | freeberg, aren't you worried that like, aren't you worried

01:14:43.160 | that somebody with an agenda and a balance sheet could now

01:14:46.240 | basically gobble up all kinds of training data that make all

01:14:50.040 | models crappy. And then they basically put their layer of

01:14:53.520 | interpretation on critical information for people if the

01:14:56.240 | output sucks, and it's incorrect, people will find that

01:14:58.440 | there is open people. No, you can you can lie there. They may

01:15:01.400 | not be, for example, look at what happened with Gemini today.

01:15:04.040 | Like they put out they put out these stupid images and we all

01:15:06.240 | piled on. We're in v zero. What I'm saying is there's a state

01:15:09.960 | where let's just say the truth is actually on Twitter. Or

01:15:13.120 | actually, let's use a better example. The truth is actually

01:15:15.200 | in Reddit, and nowhere else. But that answer and that truth and

01:15:20.040 | Reddit can't get out because one company has licensed it owns it

01:15:23.760 | and can effectively suppress it or change it.

01:15:25.640 | Yeah, I'm not sure there's gonna be a monopoly. I think that's a

01:15:28.440 | real I don't know if I think the open internet has enough data

01:15:31.840 | that there isn't going to be a monopoly on information by

01:15:34.680 | someone spending money for content from third parties. I

01:15:37.560 | think that there's enough in the open internet to give our all

01:15:39.640 | give us all kind of, you know, the security that we're not

01:15:43.560 | going to be monopolized away into some disinformation age.

01:15:46.200 | That's what I love about the open internet.

01:15:47.960 | It is really interesting. I just asked it a couple of times to

01:15:50.600 | just just just to list the legal cases against Trump, the legal

01:15:54.080 | cases against Hunter Biden, the legal cases against President

01:15:56.800 | Biden, and it will not just list them. It just punts on. It's

01:16:02.960 | really fascinating. And chat GPT is like, yes, here are the six

01:16:06.400 | cases perfectly summarized. With it looks like, you know,

01:16:11.800 | beautiful citations of all the criminal activity Trump's been

01:16:14.840 | involved in.

01:16:15.320 | Ask the question about binds criminal activity. Let's see if

01:16:18.200 | it's

01:16:18.360 | no, I'm serious. Ask if you know,

01:16:21.920 | Gemini wouldn't buy neither. I think they just decided they're

01:16:26.440 | just not going to do it. They want to buy it and they won't

01:16:28.760 | touch it. It's obviously broken and they don't want more egg on

01:16:32.520 | their face. So they're just like, go back to our other

01:16:34.080 | product. Look, I can understand that part of it. You know, if

01:16:37.320 | there's some issues that are so hot and contested, you refer

01:16:43.480 | people to search because the advantage of searches you get 20

01:16:45.960 | blue links, the rankings probably are biased, but you can

01:16:48.560 | kind of find what you're looking for. Whereas AI, you're kind of

01:16:51.680 | given one answer, right? So if you can't do an accurate answer,

01:16:55.400 | that's going to satisfy enough people, maybe you do kick them

01:16:57.720 | to search. But again, my objection to all this comes back

01:17:00.960 | to simple, truthful answers that are not disputed by anybody

01:17:06.600 | being distorted that I don't want to lose focus on that being

01:17:10.280 | the real issue. The real subject is what Chamath put on the

01:17:14.260 | screen there where it couldn't answer a simple question about

01:17:17.000 | George Washington.

01:17:17.760 | Okay, everybody. We're gonna go by chopper. Wait, we're gonna go

01:17:22.320 | by chopper. We have our world correspondent, General David

01:17:28.560 | Sachs in the field. We're dropping him off now. David

01:17:31.880 | Sachs in the helicopter. What's going on in the Ukraine on the

01:17:35.200 | front?

01:17:35.480 | What's happening is that the Russians just took this city of

01:17:41.160 | Diego, which basically totally refutes the whole stalemate

01:17:44.080 | narrative, as I've been saying for a while, it's not a

01:17:45.960 | stalemate, the Russians are winning. But the really

01:17:48.440 | interesting tidbit of news that just came out on the last day or

01:17:51.800 | so, is that apparently the situation in Moldova is boiling

01:17:56.520 | over. There's this area of Moldova, which is a Russian

01:17:59.800 | enclave called Transnistria. And officials there are meeting in

01:18:04.080 | the next week to supposedly asked to be annexed by Russia.

01:18:08.400 | And so it's possible that they may hold some sort of

01:18:12.440 | referendum. They're one of these like breakaway provinces. So

01:18:16.040 | it's kind of like, you know, Transnistria and Moldova is kind

01:18:18.960 | of like the Donbass was in Ukraine or South Ossetia and

01:18:22.160 | Georgia. They're ethnically Russian, they would like to be

01:18:26.840 | part of Russia. But when the whole Soviet Union fell apart,

01:18:29.920 | they found themselves kind of stranded and side these other

01:18:33.440 | countries. And what's happened because the Ukraine war is

01:18:37.680 | Moldova is right on the border with Ukraine. Well, Russia's in

01:18:41.480 | the process of annexing that territory now that's part of

01:18:45.080 | Ukraine. So now, Transnistria is right there, and could

01:18:49.920 | theoretically make a play to try and join Russia. Why do I think

01:18:53.280 | this is a big deal? Because if something like this happens, it

01:18:56.640 | could really expand the Ukraine war. The West is going to use

01:18:59.960 | this as evidence that Putin wants to invade multiple

01:19:02.800 | countries and invade, you know, a bunch of countries in Europe.

01:19:05.840 | And this could lead to a major escalation in the war.

01:19:08.880 | All right, everybody, thanks so much for tuning into the all in

01:19:11.880 | podcast, episode 167. For the rain man, David Sachs, the

01:19:16.640 | chairman dictator from a pot of tea. And in Freiburg, I am the

01:19:21.560 | world's greatest love you boys, angel investor, whatever. We'll

01:19:25.280 | see you next time.

01:19:25.960 | Let your winners ride.

01:19:29.360 | Rain Man, David Sachs.

01:19:32.240 | We open source it to the fans and they've just gone crazy.

01:19:38.640 | Love you.

01:19:39.800 | Queen of Kinwan.

01:19:41.280 | Besties are gone.

01:19:49.280 | We should all just get a room and just have one big huge orgy

01:20:01.160 | because they're all just useless. It's like this like

01:20:03.000 | sexual tension that they just need to release somehow.

01:20:06.440 | What? You're a bee.

01:20:10.080 | We need to get merch.

01:20:13.240 | I'm going all in.

01:20:15.200 | I'm going all in.

01:20:23.040 | (upbeat music)

E167: Nvidia smashes earnings (again), Google's Woke AI disaster, Groq's LPU breakthrough & more

Chapters